Conditional vector quantized variational autoencoders for machine sound anomaly detection under domain shift.

Area of Science:

• Industrial Engineering
• Machine Learning
• Acoustics

Background:

• Effective monitoring of industrial machinery is vital for safety and efficiency.
• Domain shift in machine sounds due to varying operational conditions hinders accurate anomaly detection.

Purpose of the Study:

• To develop an unsupervised anomaly detection method for machine sounds robust to domain shifts.
• To enhance the generalization capabilities of machine sound anomaly detection models in industrial settings.

Main Methods:

• Integration of Vector Quantized Conditional Variational Autoencoder (VQ-CVAE) with PixelSNAIL for unsupervised anomaly detection.
• Utilizing audio-text contrastive learning to extract conditional textual features for multimodal representation.
• Joint anomaly scoring in time-frequency and latent spaces using VQ-CVAE codebook and PixelSNAIL prediction loss.

Main Results:

• The proposed method demonstrates superior performance across diverse machine types on the DCASE2024 Task 2 dataset.
• Effective mitigation of feature shifts and enhanced noise robustness achieved through VQ-CVAE and residual modules.
• Refined anomaly detection results by leveraging PixelSNAIL's prediction loss in the latent space.

Conclusions:

• The developed unsupervised method is effective and robust for machine sound anomaly detection in complex industrial environments.
• The fusion of audio-text multimodal representations significantly improves model generalization under domain shift conditions.